Abstract: Math word problems (MWPs) convert natural math corpus into structured equation forms. Data sparsity is one of the main obstacles for math word understanding problem due to the high cost of human annotation efforts. However, existing work mainly start from the supervised learning perspective, making the low-resource scenario under explored. In this paper, we are the first to incorporate semi-supervised learning (SSL) framework into MWPs. We propose an uncertainty-aware unlabeled data selection strategies, which can access to reliable samples and increase the model capacity gradually. Besides, to improve the quality of pseudo equations, we incorporate two indirect supervision signals considering the semantic consistency property and grammar format constraints of generated equations. Experimental results on two benchmark MWPs datasets across different ratio of unlabeled data verify the effectiveness and generalization ability of our proposed method.
0 Replies
Loading